Numerical Uncertainty of Convolutional Neural Networks Inference for Structural Brain MRI Analysis

TL;DR

This study evaluates the numerical uncertainty in CNN-based brain MRI analysis, demonstrating that CNN predictions are more accurate and reproducible than traditional methods, using stochastic arithmetic techniques.

Contribution

It introduces the application of Random Rounding to assess numerical uncertainty in CNN inference for brain MRI analysis, highlighting improved accuracy and reproducibility.

Findings

CNN predictions are more numerically accurate than traditional methods.

CNN results show higher reproducibility across environments.

CNN models outperform traditional image-processing in accuracy metrics.

Abstract

This paper investigates the numerical uncertainty of Convolutional Neural Networks (CNNs) inference for structural brain MRI analysis. It applies Random Rounding -- a stochastic arithmetic technique -- to CNN models employed in non-linear registration (SynthMorph) and whole-brain segmentation (FastSurfer), and compares the resulting numerical uncertainty to the one measured in a reference image-processing pipeline (FreeSurfer recon-all). Results obtained on 32 representative subjects show that CNN predictions are substantially more accurate numerically than traditional image-processing results (non-linear registration: 19 vs 13 significant bits on average; whole-brain segmentation: 0.99 vs 0.92 S{\o}rensen-Dice score on average), which suggests a better reproducibility of CNN results across execution environments.